Ahiru

Maintaining a model of the world

Sun, 14 Dec 2025 15:05:20 +0200

If I look at the sky and think that it’s probably going to rain, this means something like “there are lots of dark clouds - in the past, this often resulted in rain. When there weren’t clouds, it didn’t usually rain. So it seems worth preparing for rain”. I’m not sure it will rain, I’m not sure it won’t rain - I think either might happen, but if I had to choose only one, I’d go with it going to rain? Maybe?

Binary distribution probabilities

Sat, 06 Dec 2025 23:27:03 +0200

Simple binary distribution

You often go out for dinner with your friend Joe. You both like different places, so to decide where to go, you both propose a place, then you flip a coin to choose which option. Joe is very superstitious, so you always use the same coin, which he got from his grandfather. Your method is that Joe flips, you choose. The problem with this is that you remember his grandfather once talking about a shady coin that comes out tails 60% of the time. Should you prefer heads or tails?

Gitea

Thu, 29 May 2025 12:23:09 +0100

What to do?

I want to collect notes in a more organised way. Or at all, really. Otherwise things disappear. I’ve heard good things about Obsidian so thought I’d give it a try. But I’d like to have backups of my notes and have the option to collect notes on multiple devices such that they’d automatically be synched.

Obsidian has encrypted synch support, but only if you pay. Not that it’s a lot ($4/month), but that also requires setting up payments, which involves giving out personal data and that I don’t like.

Averages and sample sizes

Fri, 27 Oct 2023 14:58:14 +0200

My brain seems to not get statistics and probability. So I was wondering about how averaging things out is influenced by the amount of items being averaged, and in particular, how the underlying distribution will change things. Here are some stories to illustrate the issue:

Minting

You’re a Dutch banker working for the famous Wisselbank. Your job is to check whether the coins coming in are good. You know that each coin should weigh 25g ± 0.5g (I’m taking a lot of liberties with this story). Your method is to count out 100 coins, and see if they weigh 2.5kg. You just got a batch that weighs 2480g, which is 20g less than expected. But since each coin is ± 0.5g, and you have 100 coins, that should be ± 50g, right?

The purpose of the law

Mon, 04 Sep 2023 20:30:09 +0200

Approaches to Ethics

There are three basic approaches to ethics, each focusing on a different aspect:

Name	Focus	Example
Virtue Ethics	Integrity	Greek heros
Deontology	Rules	Kant
Consequentialism	Outcomes	“The ends justify the means”

Most schools of ethics are a mixture of the three, but you can find some examples of purish versions, e.g. Stoicism is a pretty pure form of Virtue Ethics, Kant’s categorical imperative an example of Deontology, and Utilitarism an example of Consequentialism.

Kernel

Fri, 23 Jun 2023 15:22:43 +0200

Glossary

Basic objects

Variable

A symbol to be evaluated

Combination

A pair to be evaluated. Basically (a . b)?

Operator

The unevaluated car of a combination.

So pretty much the function of an s-expr?

Operand tree

The unevaluated cdr of a combination

Operands

The operand tree is usually a list, e.g. (+ 1 2 3 4), in which case any element of that list is an operand

Arguments

The results of evaluating operands - this is the usual case, where the evaluated values are used, rather than the operands themselves.

CryptPad

Thu, 22 Jun 2023 14:27:09 +0200

I decided that it would be nice to have a secure(ish) method of sharing docs, so went by a friends recommendation to try CryptPad. This is a sort of google docs lookalike with E2E encryption and can be hosted locally.

CryptPad has decentish docs, and the process is explained here. There were a couple of things that I had to work out on my own:

Node setup

Install NVM - this requires either logging in/out, or running the magical commands that result of it
Setup the recommended version of node: nvm install v16.14.2 && nvm use v16.14.2
Install all dependancies: npm install
Setup bower: npm install -g bower && bower install && bower update

Config

Next the config has to be, well, configured. This is described in the docs and quite obvious. There were a couple of extra things that needed to be added, but weren’t in the sample doc - but I found these in some random web search:

Balcony boxes - update

Fri, 02 Jun 2023 22:39:13 +0200

My balcony boxes turned out to be lovely. It’s not a full garden, and it by no means could support me, but it’s lush and green which is pretty much all I could expect from 2 square meters.

The top boxes are mainly a flower mix, which look like random meadow flowers such as can be found in Poland. I also have some wild strawberries, garlic and lamb’s lettuce in the top boxes, both of which are thriving (the first strawberry was eaten today).

Potential natural vegetation of Poland

Wed, 31 May 2023 12:08:19 +0200

The Białowieża Geobotanical Station has a wonderful map of the potential natural vegetation of Poland. I looked around to see if it can be bought somewhere, but the only version I could find was as an album of 12 partial maps. Not quite what I had in mind. Fortunately, the authors have made the map available over the internet as 16 bitmap images. The obvious next step was to manually join them into one giant map, which can be found here

2023 Winter Anime Season

Sun, 26 Mar 2023 21:08:19 +0200

These are the series I watched (or at least started) and my impressions. They’re more or less ordered from most enjoyable to least. I added my scores to each title (on a scale of 1 to 5). Spoilers alert, etc.

The Way of the Househusband 2

Score: 5
日本語: “Gokushufudou 2”

This is nice wholesome entertainment. The first season was better (of course), but this is still excellent. I love the juxtaposition of Yakuza in trivial, everyday situations. It’s the absurdity of it all, combined with the absolute seriousness with which they handle each situation

AGISF - Agent foundations or AI governance

Tue, 14 Mar 2023 13:02:19 +0200

Agent foundations

Embedded agents, part 1 (Demski and Garrabrant, 2018)

Everything Everywhere All at Once

Mon, 13 Mar 2023 22:54:15 +0200

Spoiler alert: Since this is about a film, I’m obviously going to reference specifics

Everything Everywhere All at Once (which I’ll abbreviate to EEAaO) won 7 Oscars yesterday, so it seemed like something that might be worth watching. The acting Oscars seem valid. The main characters really did act well. I have no idea what the editing one is given for, so that might be correct. But if that was the best screenplay and motion picture of 2022, then it seems like a really bad year for films.

AGISF - Interpretability

Tue, 07 Mar 2023 15:58:42 +0200

Week 6 of the AI alignment curriculum. Interpretability is the study of ways to, well, interpret AI models, currently mainly NNs.

Mechanistic interpretability

This aims to understand networks on the level of individual neurons.

Zoom In: an introduction to circuits (Olah et al., 2020)

Claims

Features are the fundamental unit of neural networks. They correspond to directions. These features can be rigorously studied and understood.
Features are connected by weights, forming circuits. These circuits can also be rigorously studied and understood
Analogous features and circuits form across models and tasks.

Features

Detect a single thing
Earlier layers find generic things, like gradients or curves
Later layers find specific things, like wheels or faces
Controversial whether a single neuron can detect a single thing
Polysemantic neurons respond to multiple features

Circuits

Detects more complicated things, like all orientations of a curve
Superposition -> pure features that successfully can match a complicated thing get connected to other ones (i.e. it creates later polysemantic neurons), making use of the fact that specific things tend to not go together (e.g. a picture is either of a car or a dog)

Universability

NNs tend to generate the same base detection blocks, which would suggest that later layers might also do so

Toy models of superposition (Elhage et al., 2022)

Introduces 2 forces that counteract each other:

Privileged basis - some representations can be encoded as a single neuron
Superposition - linear representations can represent more features than dimensions, which can be seen as neural networks simulating larger networks

Concept-based interpretability

Techniques for automatically probing (and potentially modifying) human-interpretable concepts stored within neural networks

The Monty Hall problem

Fri, 24 Feb 2023 11:05:24 +0200

I’m going through “All of statistics”, and one of the exercises is to work out the probabilities for the Monty Hall problem.

The basic idea is that there are 3 doors, one of which contains a prize, the others have nothing (or a goat, in the original). You choose one door, at which point Monty Hall asks if you’re sure you want that one? To make it more interesting, he chooses one of the other doors that don’t have the prize behind them and opens it. Which means that there are now only two possible doors hiding the prize.

AGISF - Task decomposition for scalable oversight

Tue, 21 Feb 2023 12:23:13 +0200

Week 4 of the AI alignment curriculum. Scalable oversight refers to methods that enable humans to oversee AI systems that are solving tasks too complicated for a single human to evaluate. Basically divide and conquer.

AI alignment landscape (Christiano, 2020)

Intent alignment -> getting AIs to want to do what you want them to do
Paul isn’t focused on reliability (how often it makes mistakes**, hopes it’ll get better along with capabilities
Well meaning != understands me. Sort of the source of most comedies of errors?
Paul also isn’t focused on how AIs will be used. AI governance is a separate can of worms
Main focus is on reducing alignment tax

Inner vs Outer

Outer alignment -> finding an objective that incentivizes aligned behavior
Inner alignment -> making sure that the policy we ended up with is robustly pursuing the objective that we used to select it

landscape graph

AGISF - Goal misgeneralization

Tue, 14 Feb 2023 17:45:19 +0200

Week 3 of the AI alignment curriculum. Goal misgeneralization is scenarios in which agents in new situations generalize to behaving in competent yet undesirable ways because of learning the wrong goals from previous training.

Goal Misgeneralisation: Why Correct Specifications Aren’t Enough For Correct Goals (Shah, 2022)

Blog post

A correct specification is needed for the learner to have the right context (so it doesn’t exploit bugs), but doesn’t automatically result in correct goals
If there are more than 1 way to interpret the set of learning examples (and there probably is), assume the worst
Good and bad behaviour can result in the same rewards - the reward function acts as a lossy compression of the value of a given set of actions
An AGI doesn’t have to be actively deceptive to have bad outcomes

Examples

Ball following

The goal is to hit the blobs in a specific sequence, the agent learns to follow the other agent, as that originally gave good results
Seems like a good heuristic? I’d probably do the same?

Tree chopping

Chopping trees as fast as possible gave the most reward when the agent wasn’t very competent, so it just chops down everything, rather than do so in a sustainable way

Expression evaluator

Always asks clarifying questions
Might not be that stupid, depends how you look at it - could be clarifying that it’s using numbers. Also that it’s showing engagement? Sounds like something that would be done in school

Paper

Capability is understood to mean that an agent is good at a task even in novel situations
Pretty much the same as the blog post, but with more details and more sciency

Why alignment could be hard with modern deep learning (Cotra, 2021)

Introduces:
- Saints - actively want to do what you want (i.e. aligned)
- Sycophants - do whatever they think you want them to do (i.e. whatever gets them praised)
- Schemers - do whatever they think you will want that will advance their hidden goals (i.e. whatever gets them closer to what they really want)
This is sort of like simulacrum levels?

Thought experiments provide a third anchor (Steinhardt, 2022)

Introduces anchors, as sort of idea generators
- Current ML systems - extrapolating from e.g. LLM can give valuable insights, and also allow testing of hypothesizes
- Humans - are an intuitive source of understanding of various concepts, seeing as they are quite intelligent and can learn
- Though experiments with ideal optimizers - useful to work out generalizations and overarching constraints
ML tends to change a lot, fast. Yesterdays models are a lot less interesting and insightful
Humans tend to anthropomorphize a lot too much, e.g. assigning emotions where there are none. Humans are useful, but very limited (like neurosurgeons vs vets)
Thought experiments are nice, but need to be checked empirically whether they make sense. They also tend to assume spherical cows
- Correctly predicted deception and power seeking
Other anchors are evolution, ethology or economy

ML systems will have weird failure modes (Steinhardt, 2022)

Discusses deceptive alignment, where the model pretends to be aligned during training, then does a treacherous turn
Requires the model to:
- have a coherent policy/plan/reward function that it wants to achieve/protect
- be able to make long term plans
- know what the desired outcomes are (so it can pretend to generate them) These are an unlikely set of assumptions (which is duely noted in the article), but show that it’s possible in theory

The alignment problem from a deep learning perspective (Ngo, Chan and Mindermann, 2022)

AGISF - Reward Misspecification and Foundation Models

Tue, 07 Feb 2023 15:01:56 +0200

Week 2 of the AI alignment curriculum. Reward misspecification occurs when RL agents are rewarded for misbehaving.

Specification gaming: the flip side of AI ingenuity (Krakovna et al., 2020)

Specification gaming is a behaviour that satisfies the literal specification of an objective without achieving the intended outcome.
Evil genies
Amounts to both the old and new understanding of hacking
RLHF can help, but only if the correct reward function is learned
The map is not the territory; agents learn from the map

Examples

Lego stacking - red block gets flipped, so its bottom is at the same height as the blue block’s top
Coast runners - the boat continuously goes for the bonuses, rather than winning the race
Grasping ball with robotic hand - the agent hovers the hand in front of the camera
Simulated robot - slides along the ground because of physics bug

Deep RL from human preferences: blog post (Christiano et al., 2017)

Might be surprising with out of distribution examples
Hidden assumptions will come out to bite you

On the opportunities and risks of foundation models (Bommasani et al., 2022)

Emergence means that the behavior of a system is implicitly induced rather than explicitly constructed; it is both the source of scientific excitement and anxiety about unanticipated consequences
Homogenization indicates the consolidation of methodologies for building machine learning systems across a wide range of applications; it provides strong leverage towards many tasks but also creates single points of failure

Learning to summarize with human feedback: blog post (Stiennon et al., 2020)

The alignment problem from a deep learning perspective (Ngo, Chan and Mindermann, 2022)

AGISF - Artificial General Intelligence

Tue, 31 Jan 2023 14:56:12 +0200

Questions

how long till models start autofeedbacking, like AlphaZero did? This suggests that it’s already happening
the general timeline seems to be AGI around 2050 - how accurate is that?

Core readings

Four Background Claims - (Soares, 2015)

General intelligence is a thing, and humans have it

The alternative view is that intelligence is just a collection of useful modules (speech, dexterity, etc.) that can be used in different contexts to solve stuff. This seems to track the nurture vs. nature debate?

Convolutional networks

Thu, 26 Jan 2023 13:55:27 +0200

How to order stuff

Vanilla NNs with fully connected layers work by each neuron receiving all the outputs of the previous layer (or the inputs in the case of input layers) and then doing calculations over them and the weights and biases of the given neuron. This works well in general, but is especially good when each input value has a specific meaning, e.g. [<number of cats>, <number of dogs>, <number of mice>, ...].

Convolution

Mon, 23 Jan 2023 14:15:52 +0200

Convolution is pretty much just sliding 2 lists along each other and summing their element wise multiplication. The general equation for the $t^{th}$ result is: $$(f*g)(t) = \int^{\infty}_{-\infty}f(\tau)g(t-\tau)d\tau$$

Properties

Convolution has a couple of interesting properties, which boil down to it being commutative:

$f * g = g * f$ - I make use of this in the convolve2() function, as it makes things easier for me to understand
$\int (f * g) = \int g \cdot \int f$. This isn’t true in general for integrals, e.g. $\int (x \cdot x) \ne \int x \cdot \int x$
$f(t) * \delta(t) = f(t)$, where $\delta(t)$ is the Dirac delta function. In the case of lists, this is convolving $f(x)$ by $[1]$, which will simply return the given element of the list.

Code version

To play about with the idea, I wrote the following implementations of convolution. The first version is pretty much a straight forward implementation of the convolution equation, while the second one is more pythonic by playing around with the lists. These don’t implement proper convolution, as they only accept lists, rather generic numerical functions. That being said, a list can totally be viewed as a function that receives an integer (the index, i.e. $x$) and returns a value ($f(x)$). The at_index() function sort of does this, as it returns $0$ for any index that isn’t a valid index of the given list.

Artificial neurons

Sat, 21 Jan 2023 17:56:43 +0200

Sigmoid ($\sigma(x) = \frac {1}{1 - e^{-z}}$)

Can saturate when $z$ is large
Has a nice derivative: $\sigma’(x) = \sigma(x)(1 - \sigma(x)) $
Transforms $(-\infty; \infty) \to (0, 1)$

Tanh ($\tanh(z) = \frac {e^z - e^{-z}}{e^z + e^{-z}}$)

Is a rescaled version of the sigmoid: $\sigma(z)=\frac {1 + \tanh(\frac z 2)}{2}$
Transforms $(-\infty; \infty) \to (-1, 1)$, so is zero centered
May require normalization of outputs (or even inputs) to a prob distribution
$\tanh’(z) = 1 - \tanh^2(z)$
Is found empirically to provide only a small or no improvement in performance over sigmoid neurons
Can also saturate

Rectified Linear neuron ($\text{ReLU}(z) = \max(0, z)$)

Doesn’t saturate
Doesn’t learn if the weighted input is negative, as the gradient is then 0

Hyperparameters

Sat, 21 Jan 2023 17:17:14 +0200

Hints when starting on a new problem

Start by getting better than change results, as a baseline for improvements
Strip the problem space down to a simpler version, e.g. just learn to classify 0 and 1, rather than all the digits of MNIST
Focus on getting decent values hyperparameters one by one (e.g. $\lambda$ or $\eta$), rather than randomly jumping around hyperparameter space
Start with getting decent learning rates etc. before scaling up the number of neurons
Initially jump about by largish amounts, looking for a decent value, then fine tune
Can be useful to intertwine various hyperparameter optimizations, as they influence each other
Pay very close attention to validation accuracy

Speed up feedback loops

Use a small subset of data to start with
Experiment with a stripped down version by removing some hidden layers
Increase the frequency of monitoring, e.g. every n batches, rather than every n epochs

Number of training epochs

Use early stopping - if there was no improvement over the last 10 epochs, stop and validate (or try different hyperparameters)

Learning rate ($\eta$)

Start by finding the threshold value for $\eta$ at which the training cost immediately begins decreasing, instead of oscillating or increasing. A decent initial value is $\eta = 0.01$. If it starts decreasing right away, keep multiplying by 10 till it doesn’t. If it didn’t decrease right away, divide by 10 until it does. This will give an order of magnitude estimation for the threshold value of $\eta$
A quick value for $\eta$ is half the threshold value

Learning rate schedule

Just use ADAM?

Regularization parameter ($\lambda$)

Start with 0, until $\eta$ is set, then set to 1. Next use the validation set to find a decent value, scaling up and down by factors of 10. Then fine tune $\lambda$, after which return to fine tuning $\eta$

Mini-batch size

Relatively independent of other hyperparameters
Mainly influences the time needed to train stuff, rather than the actual training results
Find some decent values for the other hyperparameters, then try a few batch sizes, scaling $\eta$ up and down. Choose whichever size gives the best $\frac {\text{accuracy}}{\text{clock time}}$

Automated hyperparameter search

Should read up on this? Or just use whatever fast.ai gives?

Improving NNs

Sat, 21 Jan 2023 13:12:54 +0200

Notes from chapter 3 of Neural Networks and Deep Learning

Measurements of fit

From Neural Networks and Deep Learning, 4 measures of the same data to check how good the fit is:

Here accuracy on the test data plateaus around epoch 280, while the training data cost keeps going down smoothly. On the other hand, the cost for the test data starts going up around epoch 15, which is more or less the same point that the accuracy on the training data stops drastically improving. It looks like it memorized the training data around epoch 280.
The network starts overfitting around epoch 280, which simply means that it’s not generalizing anymore
The cost function is a proxy for the overall accuracy, so the test data accuracy is more important than the test data cost - hence epoch 280, rather than 15
The difference between training and test accuracy is ~18%. This is Bad, and another sign of overfitting
One of the best ways to avoid overfitting is to increase the amount of training data. Hehe
If the model doesn’t overfit on a small subset of the data, it probably won’t fit properly on a larger set? Makes sense, but might not be true

Validation data

Early stopping is training until the classification accuracy for the validation set stops improving
Early stopping can be too soon, as sometime the accuracy takes a break and then starts improving again (something to do with grokking?)
Training data is used to check whether the parameters are getting properly trained, validation data to check whether the hyperparameters are correct
The hold out method is keeping the validation data apart from the test data, in order to find good hyperparameters

Regularization

Techniques that are used to calibrate machine learning models in order to minimize the adjusted loss function and prevent overfitting or underfitting

Cost function regularization

Works by adding an extra cost based on the weights of the NN
Basically amounts to $C = C_0 + \lambda R$, where $C_0$ is the base cost function (e.g. cross-entropy), $\lambda$ is a scaling factor called the regularization term, and $R$ is the actual regularization function

L2 regularization ($R = \sum_w w^2$)

Encourages the weights to stay small
Regularization doesn’t include the biases - only the weights
$\frac {\partial C}{\partial w} = \frac {\partial C_0}{\partial w} + \frac {\lambda}{n}w$, which is nice and simple to calculate
The weight update is $w = \begin{pmatrix} 1 - \frac {\eta \lambda}{n} \end{pmatrix}w - \eta \frac {\partial C_0}{\partial w}$. This will cause all weights to be scaled down, unless they are actively contributing to the network?
Regularizing the biases doesn’t seem to give too much, empirically. One the other hand, allowing large biases give more flexibility

L1 regularization ($R = \sum_w |w|$)

Won’t penalize outliers as much as L2 does
The weight update is $w = w - \frac {\eta \lambda}{n} \text{sign}(w) - \eta \frac {\partial C_0}{\partial w}$ (where $\text{sign}(x)$ is the sign of x). This will always shrink the weights towards 0 by a constant, as opposed to L2 where it shrunk them by an amount proportional to $w$
L1 will concentrate the weights in a small number of high importance connections, shrinking the rest to 0
Would this mean that weights don’t go to 0, just jump around it?

Network regularization

Works by directly changing the weights

Dropout

Works by temporarily removing a portion of neurons from the hidden layer for each batch
The fraction is e.g. half of the neurons
Input and output layers can’t be removed, as that would change the function signature
Intuitively, this amounts to training a large amount of smaller networks and averaging their outputs
The learned weights need to be scaled down, as (for $p=0.5$) they will be twice as large as they should be
Especially useful in large, deep networks
The network learns to manage even if some of its inputs are missing

Artificially expanding the data set

More data $\to$ better learning, but getting data is hard
Slightly rotating, expanding etc. images doesn’t make them less recognizable $\to$ adding random perpetuations to the training data will generate a lot of new but valid data
The data set is crucial - an outstanding, magnificent algorithm might turn out to simply be good at a given data set and be rubbish on other data

Weight initialization

A basic idea is Gaussian with mean 0 and std dev 1
a std dev of 1, with a lot of parameters will result in the overall std dev of $z$ to be quite large. Which in turn will saturate $\sigma$ activation functions. A much better idea is to initialize with a std dev of $\frac 1 {\sqrt {n_{in}}}$, which will result in the initial std dev of $z$ being around 1
Using $\frac 1 {\sqrt {n_{in}}}$ will result in the network start off a lot faster, but can also sometimes make the final result better

Cost functions

Fri, 20 Jan 2023 17:47:12 +0200

Quadratic cost function ($C = \frac {(y-a)^2}{2}$)

This is nice and simple, with the additional bonus that $C’ = y - a$
The loss grows exponentially, so a large error is treated a lot more harshly than a small error - this seems a good idea
It’s very xenophobic, in that it will go out of its way to harm outliers
The exponent will cause the loss to always be positive. Would allowing it to go negative result in finding optima faster? Or would it just overshoot in the opposite direction?
Bad behaviour gets punished, and good ignored. Yet it works. Beware the naturalist fallacy

Cross-entropy cost function ($C = -\sum_iy_i\log(a_i)$)

Each possible outcome will have it’s own clause. In the case of a binary classifier, this will be $-(y\log(a) + (1 - y)\log(1 - y))$, where one clause will disappear (seeing as $y \in {0, 1}$)
For tasks with $n$ categories, there will be $n$ clauses, e.g. $-(1\cdot\log(0.2) + 0\cdot\log(0.3) + 0\cdot\log(0.5))$ for the case of $y = cat1, a=[0.2, 0.3, 0.5]$
Cross-entropy works on probability distributions. So it needs inputs such that $\sum x =1, x \in <0, 1>$
The sum will be negative, since all the $a_i < 1$, so the $/log(a_i) < 0$. The minus makes sure that the loss is always non negative
During backpropagation, it’s very simply to calculate $C’$ for L2, but that leaves the activation function derivative, which in the case of it being the sigmoid function, can cause neurons to saturate. The derivative of cross-entropy cancels this out, leaving $\frac {\partial C}{\partial w_j} = x_j(a-y)$. This doesn’t saturate the output neurons (but hidden ones still can).
Was found by starting with the desired shape of $\frac {\partial C}{\partial w_j}$ and integrating backwards
Measures how surprised one should be with the difference between $y$ and $a^L$

Softmax ($a = \frac {e^{x_j}} {\sum_k e^{z_k} }$)

Transforms input vectors into probability distributions
Will work well with cross-entropy
Can be used directly to assign probabilities for each category
The exponentials ensure that all values are positive
The derivative of softmax uses some dark magic, but the result is that $\frac {\partial a_j}{\partial z_k} > 0$ for $j = k$ and $\frac {\partial a_j}{\partial z_j} < 0$ for $j \neq k$. This means that changing $z_j$ will cause changes in the opposite direction for all other $z$s
Softmax layers are nonlocal - each neuron depends on all other neurons in the layer for its outputs

Log likelihood cost function ($C = - \ln a^L_y$)

This is just the cross-entropy for multiple classes, where $y \in {0, 1}$. So for $n$ classes, where $k$ is the correct class: $$\begin{cases} y_i = 0 \text{ for } i\ne k \\\ y_i = 1 \text{ for } i = k \end{cases}$$ This ends up being $- \delta_{jy}\ln a^L_j$, because only the index for the correct class will be non zero.
Softmax + log likelihood are equivalent to sigmoid + cross-entropy. This I still don’t understand - it’s something to do with the dark magic of the derivatives.

Backpropagation

Thu, 19 Jan 2023 12:40:35 +0200

The basic equations

Given the following notation:

$w_{jk}^l$ are the weights from the $k^{th}$ neuron in the $l^{n-1}$ layer to the $j^{th}$ neuron of the $l^{th}$ layer
$a_j^l$ is the activation of the $j^{th}$ neuron in the $l^{th}$ layer
$b_j^l$ is the bias of the $j^{th}$ neuron in the $l^{th}$ layer
$z^l$ is $a^{l-1}w^l + b^l$ for the $l^{th}$ layer
$z_j^l$ is $\sum_ja_j^{l-1}w_{jk}^l + b_j^l$ for the $j^{th}$ neuron in the $l^{th}$ layer
$\sigma$ is the activation function used
$C$ is the loss of for a given input
$\frac {\partial C}{\partial a_j^L}$ is the cost function gradient for $a_j^L$
$\delta^l_j$ is the error of the $j^{th}$ neuron of the $l^{th}$ layer
$x^L$ is $x$ from the output layer, e.g. $a^L$ are the activations of the output layer

Backpropagation can be condensed into the following 4 equations:

Study plan

Thu, 17 Nov 2022 17:33:19 +0200

I want to be stronger. This is true in general, but in this case specifically I want to save the world and so need the resources to do so. The current main candidate for X-risk seems to be AI, so it would be rational to power up in order to help in that epic battle. So this is a list of skills that would be useful, along with sources from which to learn more. Sort of a skill tree?

Miracles and why not to believe them

Wed, 16 Nov 2022 10:57:24 +0200

The teachings of Jaynes-sensei

I’ve recently been trying to get through Probability Theory: The Logic of Science, which has come highly recommended, and with which opinion I so far wholeheartedly concur. It generally explains why Bayes was right, starting with a couple of basic desiderata, i.e.:

degrees of Plausibility are represented by real numbers
qualitative Correspondence with common sense
that it always reasons consistently

It turns out that these are totally sufficient to come up with a system that encompasses both probability and logic (as a binary subset of probability). From these basic assumptions, Jaynes proceeds to build up the known rules of probability, to thoroughly dis frequentialists, and to go on to show that Bayes’ rule is pretty much all you need for most scientific work, as long as you actually think through all the appropriate implications. It assumes that the reader is fluent in mathematics (i.e. basic calculus and algebra), but I personally was able to follow virtually all the proofs, so it can’t require that much.

Balcony boxes

Wed, 12 Oct 2022 23:19:51 +0200

Plants are pleasant. Growing ones own is good. Problem is that they require light and space to grow. I considered setting up a hydroponics system or - even better - aquaponics, but was informed in no uncertain terms that Not In This Flat. Luckily a large (for a flat), southern facing balcony came to the rescue. So I got me some balcony boxes to plant stuff in.

I find the idea of self-sufficiency very alluring, so I want to be able to grow as much of my own veg as possible. That means multiple large boxes. Luckily I know someone who sometimes works as a carpenter, so I got him to make me proper wooden boxes, which I’d then simply screw together. I settled on the following plan, with an extra 3, 16cm heigh boxes to go above the 3 pink ones. So 9 boxes in total:

Death and love

Thu, 06 Oct 2022 11:41:19 +0200

Child mortality

Historical rates of child mortality are abysmal. Before 1900 almost half the people born died before they were 15. Almost a third of them died before they were one year old:

This has gotten a lot better lately, with now “only” 4% of children dying before they are five years old. In developed countries it’s gotten so well that while I’m guessing that most people know someone who lost a child family member, they might not have known the child. Off the top of my mind I can think of 3 such families that I know - one child died when I was an infant (so I didn’t know her), one died soon after being born (I heard from/about the parents, but didn’t see the child) and one died when she was 4 (whom I met). While each such event is a tragedy, it’s frankly amazing that it happens so seldom! I wonder how many people living 300 years ago would believe it?

MDP

Tue, 07 Jun 2022 13:16:19 +0200

I got a new work laptop with various spyware installed. I therefore want to limit its access to my systems, what with me being paranoid. Previously, if I wanted to listen to music, I’d connect my NFS music partition and play it from my work computer. But why go through the intermediary, when I can play stuff directly from my Pi? Especially as it was originally supposed to be a media center. Seeing as my Pi is headless, I need an interface of some kind. Preferably one that can be accessed from multiple devices. The simplest option, of course, is some kind of web thingy. Especially as I already have that set up for books and torrents.

Brave

Wed, 13 Apr 2022 23:20:19 +0200

I just finished watching Brave. I was ordered to watch it, as all Disney films Must Be Known, but had been putting it off for a while. It’s pretty much what one could expect of a Disney film - the princess has duties, but wants to live her own life, and over the course of the film both she learns the value of tradition, and her parents learn the value of letting teenagers have their way. Standard fare. With very pretty landscapes and scottish accents.

About

Fri, 08 Apr 2022 21:58:07 +0200

Stuff I think about which I want to have written down for posterity to appreciate the heights of my intellect etc.

I started this after getting my own domain, mainly to have ahiru.pl actually resolve to something.

There isn’t a specific topic that I focus on - I write about whatever I’ve been thinking about lately, but also use it as a place to collect good blog posts, notes, anki collections and other such stuff. My current interests are AI safety, Anime, Biology, Books, Botany, Programming and Rationalty, in alphabetical order.

Religion without God

Wed, 30 Mar 2022 22:59:09 +0200

My sister has a classification of people where most are neutral, but some are unicorns and some are magic. The idea being that unicorns are wonderful and mystical beasts which have to be protected with everything possible because of how amazing they are, while magic people are those who do so, with the additional attribute of actively shielding them from evil and healing them when injured. The idea is surprisingly apt. Ideas can also be classified this way. I am not emotional. But some ideas are so beautiful that they move me to tears. They also tend to be delicate things that are hard to keep hold of. Some people are magic in how they protect and culture (in the gardening sense) ideas, defending them from banality and perversion.

Positive impact - Positive

Thu, 17 Mar 2022 16:16:15 +0200

This is the second part of my thoughts on how I understand “Positive Impact”, where I’m wondering what “Positive” means. The “Impact” part can be found here.

Signs

A positive value is one that is larger than zero. Not all that helpful. Though at least we know it’s not negative, which is something. The question is what (and how) is being measured. I think it’s safe to assume that a positive outcome is one that the evalutor would describe as “good”, while a negative one as “bad”. If so, then it seems like there are two main big areas to ponder, i.e. what is “good” and how many dimensions of “good” can be had at once.

Positive impact - Impacts

Wed, 16 Mar 2022 00:33:45 +0200

I’ve been asked to define how I understand “Positive impact”. Which is an exceedingly good question. Also an unexpectedly hard one. Which shouldn’t be surprising, seeing as that is pretty much the main thing studied by philosophy of morality etc. So here’s my attempt at solving one of the hardest and most elusive problems of humanity in a single blog post. Trivial, I know. Just in case it wasn’t obvious, I’m not trying to create an overarching framework of morality here, just writing down my thoughts on the topic. The second part, where I pontify on morality, can be found here.

Noteworthy blog posts

Thu, 17 Feb 2022 19:31:09 +0200

I often find myself looking for a blog post that I read 2 years ago to prove some point or other, just to end by giving up. I somewhat started to address this problem by book marking them, but that tends to get spread over various browsers. Of course there are services that offer this, hehe, service, but that goes back to my paranoia. So a manual list will do for now. I don’t tend to find more than 1 or 2 posts worth saving per week, so it shouldn’t be too unwieldy.

Lectures

Sat, 12 Feb 2022 12:31:09 +0200

As a part of my ongoing quest for excellence, world domination etc., I’ve lately been educating myself on various critical topics, e.g. that ancient islanders lived in turtle shells, a morsel of information which, as anyone with an ounce of reason will readily assert, is of the utmost importance to a person aspiring to the peaks of wisdom. A part of these educational endeavors is to go through a bunch of lectures on various topics. Most of them are saved as bookmarks on my home PC, which while fine for day to day activities, somewhat limits my scholastic improvements - it’s hard to access data from my home computer when it’s turned off… A simple solution would be to use any one of a million ways to save playlists (the most common solution of course being a Youtube profile), but I have an unnatural abhorrence of losing my precious brainly info. So rather than letting some internet scum steal my info, I decided to put it in one place so they can readily access it… Anyways - below is a list of lectures that I’m watching/planning on watching. It’ll be interesting to see how many of these stay up over the years.

Temporal-Difference Learning

Fri, 04 Feb 2022 15:44:15 +0200

Temporal-Difference Learning

I’ve been watching David Silver’s Reinforcement Learning lectures. They’re really well done, and I personally found them at the right ration of new/known info. I.e. they’re challenging, but not incomprehensible. I’m not sure if I should take that as a complement or an insult… Anyway, I was struck by how well humans can be modeled by Temporal-Difference learning. Although I see from Wikipedia that I’m not very original, seeing as there is a whole section there on how it’s relevant for neuroscience… Also Scott Alexander wrote about it recently.

Arcane

Wed, 22 Dec 2021 23:32:19 +0200

I finally got round to finishing Arcane. Mark me as impressed. All round. It starts as a steampunkish dystopia and just gets better. I mean who doesn’t love magic in their gear sets? The diversity of characters is astounding. Which is of course to be expected from a series based on LOL, but I’m still impressed they managed to pull it off. And it’s not just the characters which are diverse. The series starts out in the undercity (the name is totally apt), but soon proceeds to show the overcity (proper steampunk), and also has other “environments” to show.

Gatunki roślin

Wed, 15 Dec 2021 21:27:09 +0200

Fiszki z roślin Polski

Chcę wiedzieć jaki rośliny spotykam. Ale mam słabą pamięć. Na ratunek przychodzi Anki. Tutaj jest zbiór fiszek co by z tym wspomóc.

Łacinska nazwa	Nazwa pospolita
Achillea filipendulina	Krwawnik wiązówkowaty
Achillea millefolium	Krwawnik pospolity
Aconitum firmum	Tojad mocny
Actaea spicata	Czerniec gronkowy
Adenostyles alliariae	Miłosna górska
Aegopodium podagraria	Podagrycznik pospolity
Ajuga reptans	Dąbrówka rozłogowa
Alliaria petiolata	Czosnaczek pospolity
Allium ursinum	Czosnek niedźwiedzi
Alnus glutinosa	Olsza czarna
Alnus incana	Olsza szara
Anagallis arvensis	Kurzyślad polny
Anemone nemorosa	Zawilec gajowy
Archangelica officinalis	Arcydzięgiel litwor
Arctium lappa	Łopian większy
Arctium minus	Łopian mniejszy
Armoracia rusticana	Chrzan pospolity
Arnica montana	Arnika górska
Artemisia absinthium	Bylica piołun
Artemisia vulgaris	Bylica pospolita
Asarum europaeum	Kopytnik pospolity
Atropa belladonna	Pokrzyk wilcza jagoda
Bellis perennis	Stokrotka pospolita
Berberis vulgaris	Berberys zwyczajny
Betula pendula	Brzoza brodawkowata
Betula pubescensBidens	Brzoza omszona
Bidens tripartita	Uczep trójlistkowy
Calystegia sepium	Kielisznik zaroślowy
Campanula trachelium	Dzwonek pokrzywolistny
Capsella bursa-pastoris	Tasznik pospolity
Cardamine pratensis	Rzeżucha łąkowa
Carlina acaulis	Dziewięćsił bezłodygowy
Carpinus betulus	Grab pospolity
Carum carvi	Kminek zwyczajny
Centaurea cyanus	Chaber bławatek
Centaurea jacea	Chaber łąkowy
Centaurium erythraea	Centuria pospolita
Cerastium arvense	Rogownica polna
Chelidonium majus	Glistnik jaskółcze ziele
Chenopodium album	Komosa biała
Cicerbita alpina	Modrzyk górski
Cichorium intybus	Cykoria podróżnik
Cicuta virosa	Szalej jadowity
Cirsium arvense	Ostrożeń polny
Cirsium oleraceum	Ostrożeń warzywny
Colchicum autumnale	Zimowit jesienny
Consolida regalis	Ostróżeczka polna
Convallaria majalis	Konwalia majowa
Convolvulus arvensis	Powój polny
Corydalis solida	Kokorycz pełna
Crocus scepusiensis	Szafran spiski
Cykoria podróżnik	Cichorium intybus
Czosnek pospolity	Allium sativum
Daphne mezereum	Wawrzynek wilczełyko
Daucus carota	Marchew zwyczajna
Dianthus deltoides	Goździk kropkowany
Doronicum clusii	Omieg kozłowiec
Dryas octopetala	Dębik ośmiopłatkowy
Echinops sphaerocephalus	Przegorzan kulisty
Echium vulgare	Żmijowiec zwyczajny
Elymus repens	Perz właściwy
Epilobium angustifolium	Wierzbówka kiprzyca
Epipactis helleborine	Kruszczyk szerokolistny
Eryngium planum	Mikołajek płaskolistny
Erysimum cheiranthoides	Pszonak drobnokwiatowy
Eupatorium cannabinum	Sadziec konopiasty
Euphorbia cyparissias	Wilczomlecz sosnka
Fagus sylvatica	Buk zwyczajny
Ficaria verna	Ziarnopłon wiosenny
Fragaria vesca	Poziomka pospolita
Fumaria officinalis	Dymnica pospolita
Galanthus nivalis	Śnieżyczka przebiśnieg
Galeobdolon luteum	Gajowiec żółty
Galeopsis tetrahit	Poziewnik szorstki
Galinsoga quadriradiata	Żółtlica owłosiona
Galium odoratum	Przytulia wonna, marzanka wonna
Gentiana asclepiadea	Goryczka trojeściowa
Geranium pratense	Bodziszek łąkowy
Geranium pyrenaicum	Bodziszek pirenejski
Geranium robertianum	bodziszek cuchnący
Glechoma hederacea	Bluszczyk kurdybanek
Heracleum sphondylium	Barszcz zwyczajny
Hieracium pilosella	Jastrzębiec kosmaczek
Humulus lupulus	Chmiel zwyczajny
Hyoscyamus niger	Lulek czarny
Hypericum perforatum	Dziurawiec zwyczajny
Impatiens glandulifera	Niecierpek gruczołowaty
Impatiens noli-tangere	Niecierpek pospolity
Iris pseudacorusJuglans	Regia Orzech włoski
Lamium purpureumLaserpitium	Prutenicum Okrzyn łąkowy
Lathraea squamaria	Łuskiewnik różowy
Lathyrus pratensis	Groszek żółty
Leontopodium nivale	Szarotka alpejska
Leucanthemum vulgare	Złocień właściwy
Leucojum vernum	Śnieżyca wiosenna
Linaria vulgarisLithospermum	Officinale nawrot lekarski
Lotus corniculatus	Komonica zwyczajna
Lycopus europaeus	Karbieniec pospolity
Lysimachia vulgaris	Tojeść pospolita
Lythrum salicaria	Krwawnica pospolita
Malva neglecta	Ślaz zaniedbany
Melilotus officinalis	Nostrzyk żółty
Menyanthes trifoliata	Bobrek trójlistny
Myosotis sylvatica	Niezapominajka leśna
Nerium oleander	Oleander pospolity
Nymphaea alba	Grzybienie białe
Odontites vernusOenothera	Biennis Wiesiołek dwuletni
Oxalis acetosella	Szczawik zajęczy
Papaver rhoeas	Mak polny
Paris quadrifolia	Czworolist pospolity
Pastinaca sativa	Pasternak zwyczajny
Persicaria amphibia	Rdest ziemnowodny
Petasites hybridus	Lepiężnik różowy
Pimpinella anisum	Biedrzeniec anyż
Pinguicula alpina	Tłustosz alpejski
Plantago lanceolata	Babka lancetowata
Plantago major	Babka zwyczajna
Polygonum persicaria	Rdest plamisty
Potentilla anserina	Pięciornik gęsi
Potentilla erecta	Pięciornik kurze ziele
Primula elatior	Pierwiosnek wyniosły
Primula veris	Pierwiosnka lekarska
Prunella vulgaris	Głowienka pospolita
Quercus petraea	Dąb bezszypułkowy
Quercus robur	Dąb szypułkowy
Ranunculus acrisRanunculus	Sceleratus Jaskier jadowity
Reseda lutea	Rezeda żółta
Ricinus communisRubus	Caesius Jeżyna popielica
Rumex acetosa	Szczaw zwyczajny
Saponaria officinalis	Mydlnica lekarska
Saxifraga aizoides	Skalnica nakrapiana
Sedum acre	Rozchodnik ostry
Silene latifolia	Lepnica biała
Sinapis arvensis	Gorczyca polna
Solanum dulcamara	Psianka słodkogórz
Solidago gigantea	Nawłoć późna
Solidago virgaurea	Nawłoć pospolita
Sonchus oleraceus	Mlecz zwyczajny
Stachys officinalis	Bukwica lekarska
Stachys palustris	Czyściec błotny
Succisa pratensis	Czarcikęs łąkowy
Symphytum officinale	Żywokost lekarski
Tanacetum vulgare	Wrotycz pospolity
Taraxacum officinale	Mniszek lekarski
Taxus baccata	Cis pospolity
Thlaspi arvense	Tobołki polne
Tilia cordata	Lipa drobnolistna
Tilia platyphyllos	Lipa szerokolistna
Trientalis europaea	Siódmaczek leśny
Trifolium pratense	Koniczyna łąkowa
Tripleurospermum maritimum	Maruna nadmorska, maruna bezwonna
Tussilago farfara	Podbiał pospolity
Urtica dioica	Pokrzywa zwyczajna
Valeriana officinalis	Kozłek lekarski
Veronica chamaedrys	Przetacznik ożankowy
Vicia cracca	Wyka ptasia
Vicia dumetorum	Wyka zaroślowa
Vinca minor	Barwinek pospolity
Viola arvensis	Fiołek polny